USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance

USDB is ByteDance’s storage architecture for recommender systems that must handle long, detailed user histories. Its central problem is an imbalance: user events arrive as frequent, small updates, while inference and training repeatedly need to read large histories. USDB combines streaming and batch storage around shared sequence-maintenance work. CascadeLSM uses time-based compaction to produce histories tailored to applications, and differentiated replicas reuse persistent files rather than maintaining wholly separate copies. For inference, partial dumps combine recently captured state with versioned historical data to reconstruct model inputs. A hybrid columnar format supports efficient reads of selected fields, while aggregation pushdown reduces the data clients must receive.

The system’s key argument is that storage work can be reused across ingestion, serving, and training, improving both speed and resource efficiency. Production migration records report an 85.8% reduction in mean write latency and 42.1% lower infrastructure cost under normalized accounting. In controlled tests against an optimized RocksDB sequence-storage baseline, USDB reads long histories faster across different field projections and client loads. When receive capacity is constrained, aggregation pushdown achieves 8.5 times the throughput of ordinary reads. Production results also show history reconstruction for serving and direct-file access for training, illustrating how a shared storage design can support fresh inference inputs alongside repeated access to historical data.